# ERNIE-Image
ERNIE-Image is an open-weights text-to-image generation model developed by the ERNIE-Image team at Baidu, built on a single-stream Diffusion Transformer (DiT) with 8 billion parameters and released in April 2026.[1](https://github.com/baidu/ERNIE-Image/blob/main/README.md) It is a separate model line from Baidu's ERNIE (Wenxin) large language models; the sources documenting it describe it only as the work of the ERNIE-Image team and do not establish any technical relationship to the [ERNIE](https://www.edgepedia.org/ernie-model-family) text models or to the ERNIE Bot consumer product.[1](https://github.com/baidu/ERNIE-Image/blob/main/README.md) This article covers the 2026 open-weights release; earlier Baidu image-generation history is not documented in the available record.
| Fact | Detail |
|---|---|
| Developer | ERNIE-Image team at Baidu[1](https://github.com/baidu/ERNIE-Image/blob/main/README.md) |
| First open-weights release | April 15, 2026[2](https://ernie-image.github.io/) |
| Architecture | Single-stream DiT, 8B parameters, latent diffusion[3](https://arxiv.org/html/2605.25347v1) |
| Components released | ERNIE-Image, ERNIE-Image-Turbo, Prompt Enhancer, ERNIE-Image-Aes, ERIA-1K benchmark[3](https://arxiv.org/html/2605.25347v1) |
| Vendor-reported GenEval | 0.89 overall without Prompt Enhancer[3](https://arxiv.org/html/2605.25347v1) |
| Vendor-reported LongText-Bench | 0.973 overall accuracy, English and Chinese[3](https://arxiv.org/html/2605.25347v1) |
| Availability | Weights hosted publicly on Hugging Face under the baidu organization[4](https://huggingface.co/baidu/ERNIE-Image) |
| Undisclosed | Training-data sources and licensing, safety evaluation, watermarking[2](https://ernie-image.github.io/) |
## What ERNIE-Image is
ERNIE-Image generates images from text prompts. Baidu describes it as reaching state-of-the-art performance among open-weight models with only 8B DiT parameters, and pairs the generator with a lightweight Prompt Enhancer that expands brief user inputs into richer structured descriptions before generation.[1](https://github.com/baidu/ERNIE-Image/blob/main/README.md) The weights are hosted publicly on [Hugging Face](https://www.edgepedia.org/hugging-face) under the baidu organization, confirming open-weights rather than API-only distribution.[4](https://huggingface.co/baidu/ERNIE-Image)
The Hugging Face Diffusers documentation, which distributes the model, describes it as a highly efficient 8B-parameter model whose performance is comparable in some scenarios to larger 20B+ models, and notes that two models had been released at the time of writing.[5](https://huggingface.co/docs/diffusers/api/pipelines/ernie_image)
## Release timeline and versions
The release on April 15, 2026 packaged two checkpoints that share an identical 8B ErnieImageTransformer2DModel backbone but differ in how they are sampled at inference time: a multi-step quality checkpoint and a distilled few-step Turbo sibling.[2](https://ernie-image.github.io/) The distilled ERNIE-Image-Turbo requires only 8 function evaluations (NFEs) to generate an image and was trained with an MT-DMD multi-expert distillation strategy, according to the technical report.[3](https://arxiv.org/html/2605.25347v1)
Beyond the two generators, Baidu released the Prompt Enhancer and ERNIE-Image-Aes, an aesthetic-scoring model, together with ERIA-1K, an aesthetic benchmark described in the technical report as carefully human-annotated.[3](https://arxiv.org/html/2605.25347v1) The available record does not document any earlier Baidu text-to-image releases or version history before this 2026 open-weights line.
## Architecture and training as published
The technical report describes ERNIE-Image as a latent diffusion model built on a single-stream DiT architecture with 8B parameters. Baidu adopted the FLUX.2 VAE as its variational autoencoder.[3](https://arxiv.org/html/2605.25347v1) Unlike other open-source models, the team used Ministral-3B as the text encoder, which they state reduces overall memory consumption and lowers the barrier to use.[3](https://arxiv.org/html/2605.25347v1)
Training followed a three-stage strategy at progressively increasing resolutions, starting at 256x256 and 512x512 before moving to higher resolutions.[3](https://arxiv.org/html/2605.25347v1) The data pipeline started from a massive raw image corpus collected from an internal data pool and progressively imposed structure through fine-grained categorization, caption enrichment with a fine-tuned vision-language model, aesthetic scoring, and hierarchical sampling.[3](https://arxiv.org/html/2605.25347v1) Post-training included diversified SFT annotations and an improved DPO alignment step, which the authors say needs only a few steps, minimizing reward hacking; this claim is vendor-reported and unverified.[3](https://arxiv.org/html/2605.25347v1)
What Baidu did not disclose is as significant as what it did. An independent analysis of the release finds that source datasets, licensing provenance, scale (number of image-text pairs), language mix, and synthetic-data fraction are not published, and that the diffusion objective, noise schedule, reward model, and reinforcement-learning algorithm behind Turbo are undocumented.[2](https://ernie-image.github.io/)
## Benchmark performance: vendor claims versus independent tests
All published benchmark numbers for ERNIE-Image come from Baidu's own technical report; no independent third-party evaluation appears in the available record.
On GenEval, Baidu reports an overall score of 0.89 for ERNIE-Image without the Prompt Enhancer, the highest in its comparison table, with particularly strong performance in spatial positioning and object-attribute coordination.[3](https://arxiv.org/html/2605.25347v1) On LongText-Bench, which measures text rendering inside images, Baidu reports an overall accuracy of 0.973 across both English and Chinese subsets, with a narrow gap between the two languages.[3](https://arxiv.org/html/2605.25347v1)
In Baidu's comparison table against Nano Banana 2.0, Wan 2.7-Image, Seedream 5.0, [Qwen-Image](https://www.edgepedia.org/qwen-image), Z-Image-Turbo, and Flux.2-Klein, the closed-source Nano Banana 2.0 leads in all categories, and Baidu states that ERNIE-Image is currently the closest open-source model, outperforming the larger Qwen-Image series with only 8B parameters.[3](https://arxiv.org/html/2605.25347v1)
An independent analysis of the release disputes the comparability of these results. It notes that the 0.973 LongText-Bench average is the single strongest benchmark result in the release, and also the only benchmark in the suite that most competing open text-to-image models do not currently report, making direct head-to-head comparison impossible.[2](https://ernie-image.github.io/) The disagreement is unresolved: Baidu's table places ERNIE-Image above Qwen-Image, while the independent analysis holds that the benchmark suite does not support that comparison.
## References
1. [baidu/ERNIE-Image README](https://github.com/baidu/ERNIE-Image/blob/main/README.md)
2. [ERNIE Image — Baidu's Open 8B DiT Text-to-Image Model](https://ernie-image.github.io/)
3. [ERNIE-Image Technical Report](https://arxiv.org/html/2605.25347v1)
4. [baidu/ERNIE-Image · Hugging Face](https://huggingface.co/baidu/ERNIE-Image)
5. [Ernie-Image · Hugging Face Diffusers docs](https://huggingface.co/docs/diffusers/api/pipelines/ernie_image)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Image generation models*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026; Oct 21, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license